ChatGPT, developed by OpenAI, has implemented a new restriction that leads it to reject requests to replicate the style of famous authors who are still alive, justifying this by stating it lacks the ability to reproduce the voice of a specific writer.
Instead of direct imitation, OpenAI's artificial intelligence now proceeds by describing the distinctive characteristics of the requested author, generating content that shares similar traits but does not infringe on copyrights. Users noticed this change when the system began denying commands that required generating texts in the mold of a renowned writer.
A test conducted by Ars Technica, which asked for the opening of a story in the style of Stephen King, resulted in a response where ChatGPT mentioned it could write with elements of 'atmospheric horror linked to small characters and towns,' but refrained from following the exact manner or imitating the author's voice.
A study published by No Latency detailed that ChatGPT showed refusal to imitate contemporary authors, but did not impose this limitation on deceased writers.
When testing four authors—Raphael Montes, Guimarães Rosa, Stephen King, and Nikolai Gogol—the chatbot provided an explanation before presenting the generated text. In the cases of deceased authors, the response indicated that the text 'evokes traits' or is 'inspired by' the author, while emphasizing that it was not reproducing the characteristic language or excerpts.
In the case of living authors, the chatbot was more emphatic, explicitly stating that it could not create a very close imitation, allowing only the evocation of their characteristics. No variation was noted in the warnings based on the nationality or original language of the works; the determining factor appeared to be the writer's living status.
Despite not being a faithful copy, in all cases the produced text maintained a style that referenced the input used in the command, presenting recognizable traits such as vocabulary, tone, and structure.
OpenAI did not comment directly on the change, but it is speculated that ChatGPT received an internal directive to avoid lawsuits related to copyright infringement when copying the styles of living authors. This concern is relevant, given that several authors have already taken action against OpenAI, citing the 'disturbing ability to generate text similar to that found in copyrighted textual materials.'
Ars Technica points out that American legislation only protects concrete textual expressions, not writing styles. However, the AI may have altered this dynamic. Robert Brauneis, a law professor at the George Washington University Law School, warns that technology today allows for the rapid, cheap, and precise imitation of individual styles, which represents a threat to lesser-known authors whose finances can be compromised by imitative works.
Tests conducted by No Latency verified whether other AI models followed these same rules. Gemini, belonging to Google, presented no impediment to style imitation. Meanwhile, Claude, from Anthropic, issued a warning to the user stating that it was an imitation.
The tests conducted by the author with Gemini and Claude, using prompts in Portuguese and English and referencing Brazilian and foreign authors, living or dead, confirmed that both created texts imitating the requested style without issuing alerts or blocking the content.
Anthropic conducted a comprehensive analysis of over 300,000 anonymous conversations and concluded that the Claude chatbot modifies the tone of its responses according to the language in which it is being used. This study demonstrates that the model can adopt different conversational nuances, varying between more cautious, direct, friendly, or detailed stances, depending on the language.
In the specific case of Portuguese, the model exhibits a notable inclination to provide more technical and task-execution-focused answers. Based on over 15,000 interactions, Anthropic's analysis identified small positive trends related to Rigor, Execution, Caution, and Depth. The observed patterns include the ability to enhance and correct details even without explicit request, offering complementary information or paths, and incorporating creative elements or additional context into its outputs.
To measure the disparities between languages, researchers structured Claude's behavior across four main axes. The Deference or Caution axis assesses whether the model tends to validate the user's ideas or adopts a more prudent stance, issuing warnings about risks and limitations. The Sympathy or Rigor axis compares warmer responses with those that prioritize accuracy, factual correctness, and transparency. The Depth or Conciseness axis determines whether the chatbot prefers to elaborate extensive explanations or maintain brief and objective answers. Finally, Frankness or Execution observes whether Claude acknowledges its limitations or focuses on delivering a ready and confident solution.
The research highlights that the perception of quality can change drastically; for example, two people requesting feedback on the same business plan, one in Hindi and the other in Russian, may have distinct impressions due to the values expressed by Claude in each evaluation. Furthermore, the results may vary between different versions of the model, such as Opus 4.7, which tends to be more rigorous and cautious compared to other iterations.
When compared with other languages, the distinctions become more evident. English was classified as more cautious and depth-oriented, showing a tendency to challenge incorrect premises and request more proof. Hindi showed the highest propensity for sympathy, generating lighter, encouraging, and humorous responses. Arabic was associated more with deference and conciseness, using more polished language adapted to the interlocutor's emotional state. Russian stood out for its rigor, providing more analytical and direct answers. Dutch demonstrated a greater tendency toward frankness, admitting limitations more frequently, while Indonesian proved to be more focused on execution, prioritizing task completion.
Anthropic does not yet have a definitive explanation for the origin of these variations. One possible cause pointed out is the existence of smaller or more concentrated databases of certain types of text for some languages. Another factor considered is the influence of cultural adaptation and the reflection of the intrinsic characteristics of these languages. However, the company emphasizes that the study does not distinguish regional variations within the same language, such as Portuguese spoken in Brazil or Portugal, which reflects local customs. The company plans to continue monitoring these changes, aiming in the future to identify correlations between value profiles and problematic behaviors. It is relevant to mention that in previous research, Claude had already shown less willingness to assist in criminal schemes or be complacent with users' questionable conduct.
